Linus Torvalds used Claude Code to build a Linux page cache fix that hit 23x read throughput, but kernel reviewers found a race before anything merged.
Or rather, LLMs helped create bad code, and so didn’t do a significant peice of work - or did it so poorly it was dangerous to rely on for an Operating system:
Race conditions in programming occur when multiple processes or threads attempt to modify and access shared data simultaneously, leading to unpredictable and unintended outcomes. They can result in data corruption, program failures, and other unexpected behaviors if not properly managed with synchronization techniques.
The work isn’t done if it’s introducing risky/bad code (especially at the level of an operating system). That’s like saying “LLMs cure cancer by cutting patients into quarters, human reviewers stop it”
Then someone claiming “the real story here is that LLMs cured cancer”.
Was cancer actually cured? The article doesn’t say the problem was fixed.
“but”? Isn’t that exactly what reviewers are for?
The actual news of the article: LLM helped do a significant piece of work that human developers had failed to do for two years.
Or rather, LLMs helped create bad code, and so didn’t do a significant peice of work - or did it so poorly it was dangerous to rely on for an Operating system:
The work isn’t done if it’s introducing risky/bad code (especially at the level of an operating system). That’s like saying “LLMs cure cancer by cutting patients into quarters, human reviewers stop it”
Then someone claiming “the real story here is that LLMs cured cancer”.
Was cancer actually cured? The article doesn’t say the problem was fixed.
The same happens when humans write code. Reviewers hopefully find some of their mistakes, then they fix them
Tell me then, was the problem solved, the work done by the LLM?
Was it considered good enough for a mainbranch?
Or was it faulty/risky and require highly trained experts to catch? Then flagged, and only put into an experimental branch?
How many ex-Oracle memory experts do you think are available for this work, in which even Linus didn’t catch the race condition?
“Hey guys, nothing to see here if we have enough high level engineers fixing everything”.
partially
Probably, if it took multiple experts to catch the bug.
As happens with human code all the time
No idea, at least one.
A competent person like Linus can manage and direct competent code generation, whether that code is generated by an AI or an intern.
Right the guy who failed to catchy the problem in the generated code, will be fine.